On 2016/9/25 8:06, Mike Kravetz wrote:
> On 09/23/2016 07:56 PM, zhong jiang wrote:
>> On 2016/9/24 1:19, Mike Kravetz wrote:
>>> On 09/22/2016 06:53 PM, zhong jiang wrote:
>>>> At present, we need to call hugetlb_fix_reserve_count when
>>>> hugetlb_unrserve_pages fails,
>>>> and PagePrivate will decide hugetlb reserves counts.
>>>>
>>>> we obtain the page from page cache. and use page both lock_page and
>>>> mutex_lock.
>>>> alloc_huge_page add page to page chace always hold lock page, then bail
>>>> out clearpageprivate
>>>> before unlock page.
>>>>
>>>> but I' m not sure it is right or I miss the points.
>>> Let me try to explain the code you suggest is unnecessary.
>>>
>>> The PagePrivate flag is used in huge page allocation/deallocation to
>>> indicate that the page was globally reserved. For example, in
>>> dequeue_huge_page_vma() there is this code:
>>>
>>> if (page) {
>>> if (avoid_reserve)
>>> break;
>>> if (!vma_has_reserves(vma, chg))
>>> break;
>>>
>>> SetPagePrivate(page);
>>> h->resv_huge_pages--;
>>> break;
>>> }
>>>
>>> and in free_huge_page():
>>>
>>> restore_reserve = PagePrivate(page);
>>> ClearPagePrivate(page);
>>> .
>>> <snip>
>>> .
>>> if (restore_reserve)
>>> h->resv_huge_pages++;
>>>
>>> This helps maintains the global huge page reserve count.
>>>
>>> In addition to the global reserve count, there are per VMA reservation
>>> structures. Unfortunately, these structures have different meanings
>>> depending on the context in which they are used.
>>>
>>> If there is a VMA reservation entry for a page, and the page has not
>>> been instantiated in the VMA this indicates there is a huge page reserved
>>> and the global resv_huge_pages count reflects that reservation. Even
>>> if a page was not reserved, a VMA reservation entry is added when a page
>>> is instantiated in the VMA.
>>>
>>> With that background, let's look at the existing code/proposed changes.
>> Clearly.
>>>> diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c
>>>> index 4ea71eb..010723b 100644
>>>> --- a/fs/hugetlbfs/inode.c
>>>> +++ b/fs/hugetlbfs/inode.c
>>>> @@ -462,14 +462,12 @@ static void remove_inode_hugepages(struct inode
>>>> *inode, loff_t lstart,
>>>> * the page, note PagePrivate which is used in case
>>>> * of error.
>>>> */
>>>> - rsv_on_error = !PagePrivate(page);
>>> This rsv_on_error flag indicates that when the huge page was allocated,
>> yes
>>> it was NOT counted against the global reserve count. So, when
>>> remove_huge_page eventually calls free_huge_page(), the global count
>>> resv_huge_pages is not incremented. So far, no problem.
>> but the page comes from the page cache. if it is. it should implement
>> ClearPageprivate(page) when lock page. This condition always true.
>>
>> The key point is why it need still check the PagePrivate(page) when page
>> from
>> page cache and hold lock.
> You are correct. My apologies for not seeing your point in the original
> post.
>
> When the huge page is added to the page cache (huge_add_to_page_cache),
> the Page Private flag will be cleared. Since this code
> (remove_inode_hugepages) will only be called for pages in the page cache,
> PagePrivate(page) will always be false.
>
> The comments in this area should be changed along with the code.
>
Thanks, I will resend the patch.
